DI
SC
US
SI
ON
P
AP
ER
S
ER
IE
S
Forschungsinstitut zur Zukunft der ArbeitInstitute for the Study of Labor
Leaving Poverty Behind? The Effects of Generous Income Support Paired with Activation
IZA DP No. 8245
June 2014
Simen MarkussenKnut Røed
Leaving Poverty Behind?
The Effects of Generous Income Support Paired with Activation
Simen Markussen Ragnar Frisch Centre for Economic Research
Knut Røed
Ragnar Frisch Centre for Economic Research and IZA
Discussion Paper No. 8245 June 2014
IZA
P.O. Box 7240 53072 Bonn
Germany
Phone: +49-228-3894-0 Fax: +49-228-3894-180
E-mail: [email protected]
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in this series may include views on policy, but the institute itself takes no institutional policy positions. The IZA research network is committed to the IZA Guiding Principles of Research Integrity. The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center and a place of communication between science, politics and business. IZA is an independent nonprofit organization supported by Deutsche Post Foundation. The center is associated with the University of Bonn and offers a stimulating research environment through its international network, workshops and conferences, data service, project support, research visits and doctoral program. IZA engages in (i) original and internationally competitive research in all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research results and concepts to the interested public. IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper should account for its provisional character. A revised version may be available directly from the author.
IZA Discussion Paper No. 8245 June 2014
ABSTRACT
Leaving Poverty Behind? The Effects of Generous Income Support Paired with Activation* We evaluate a comprehensive activation program in Norway targeted at hard-to-employ social assistance claimants with reduced work capacity. The program offers a combination of tailored rehabilitation, training and job practice, and a generous, stable, and non-means-tested benefit. Its main aims are to mitigate poverty and subsequently promote self-supporting employment. Our evaluation strategy exploits a geographically staggered program introduction, and the causal effects are identified on the basis of changes in employment prospects that coincide with local program implementation in a way that correlates with the predicted probability of becoming a participant. We find that the program raised employment prospects considerably. JEL Classification: C21, C26, H55, I30, J24 Keywords: poverty, vocational rehabilitation, social insurance, treatment effects,
program evaluation Corresponding author: Knut Røed The Ragnar Frisch Centre for Economic Research Gaustadalléen 21 0349 Oslo Norway E‐mail: [email protected]
* This paper is part of the project “Evaluation of a Norwegian Qualification Program”, financed by the Norwegian ministry of labor and social affairs. Thanks to Simen Gaure for invaluable help with estimation. Thanks also to Kristian Heggebø, Ivar Lødemel, Angelika Schaft, and seminar participants in Bergen, Stavanger, and Sydney for comments and suggestions. Data made available by Statistics Norway have been essential for the research project.
3
1 Introduction
How should policy makers design social insurance institutions in order to fight persistent un-
employment, marginalization, and poverty? While economists often emphasize the moral
hazard problems and the potential lock-in effects arising from generous social insurance pro-
grams, there is also an extensive literature focusing on the barriers associated with poverty
itself, caused, e.g., by the type of myopic behaviors it promotes and the kind of unhelpful so-
cial networks it gives access to; see Dasgupta and Ray (1986; 1987), Dasgupta (1997), Calvó-
Armengol and Jackson (2004), and Shah et al. (2012). Empirical evidence indicates that se-
vere scarcity reduces the ability to focus and concentrate on issues beyond the immediate
needs, and that it deprives sleep, erodes self-control, and reduces work productivity; see Mul-
lainathan and Shafir (2013) for a recent discussion of the literature. Some income support may
thus be required to break out of a poverty trap. But since income support is normally tested
against earned income, generosity may discourage self-sufficiency and create a benefit trap
instead. A potential solution to this dilemma is to couple generosity with activation, effective-
ly removing the leisure-component from a life on income support, and also ensuring some
“maintenance” of basic employment-skills. Properly designed, activation requirements facili-
tate more ambitious social programs without aggravating moral hazard problems; see, for ex-
ample, Moffitt (2007) for a review of empirical evidence in relation to the introduction of
activation requirements in the cash-based welfare program for single mothers in the US, and
Røed (2012) for a recent survey of the literature regarding activation strategies in unemploy-
ment and disability insurance programs.
In the present paper, we evaluate a “Qualification Program” (QP) that was launched by the
Norwegian government in 2007 as its major tool to fight poverty. The program is both costly
and ambitious, and designed to combine generosity and activation on a hitherto unprecedented
scale. It is targeted at persons with severely reduced earnings capacity and no or very limited
social insurance entitlements. The typical recruitment base is persons who have become, or
are in danger of becoming, completely reliant on means tested social assistance (welfare). QP
participants may have a variety of problems in relation to a competitive labor market, such as
poor language skills, disrupted schooling, little or no work experience, and sometimes mental
disorders and drugs problems. Around half of them have a minority background. The aim of
QP is to help these often hard-to-employ persons into self-supporting employment through an
individually tailored activation program under which they also receive a standardized (and not
means-tested) income support amounting to 170,000 NOK (approximately 28,000 $) per year
4
in 2013 (113,000 NOK for participants below 25 years) plus child allowances and housing
benefits. For the average participant, this implies a significant increase in total income, from
154,000 NOK (26,000 $) in the last whole calendar year prior to entry to 240,000 NOK
(40,000 $) in the first calendar year after entry.1 It is notable that the QP benefit is paid out by
the local municipality’s pay office rather than by the local welfare office, it is taxed like regu-
lar labor earnings and also entails the same holiday and leave privileges. The motivation for
all this is to minimize the shame and stigma associated with participating in the program, and
thus to reduce the risk of social isolation and withdrawal from networks that may be of im-
portance in the process of breaking out of poverty; see, e.g., Walker et al. (2013). Claimant
interviews have indicated that this is appreciated by the participants, and that the labor-
earnings-like payment does represent a source of dignity; see Gubrium and Lødemel (2014).
The QP benefit is normally granted for a period of up to two years, but additional extensions
can be given on the basis of an individual assessment. In return, the participant is expected to
take fully part in a tailored qualification and activation plan agreed upon by the claimant and
the caseworker. Failure to do so – without justifiable cause – immediately results in a corre-
sponding pay cut. The contents of the participation period vary a lot, but are normally made
up of a combination of consultations, employment-training, medical rehabilitation or therapy,
social training, and skills-upgrading. Large administrative resources have been set aside to
ensure that the caseworkers have the capacity needed to follow up their QP clients, and,
hence, to ensure meaningful program activities. Statistics reported by the Norwegian Labor
and Welfare Administration (NAV), indicate that each QP caseworker on average has 18 cli-
ents. By comparison, caseworkers dealing with the temporary disability insurance program on
average have 86 clients.
An important concern in the design of QP has been to achieve an appropriate balance between
individual rights and responsibilities. One could perhaps say that the basic idea of the program
is to offer income safety and stability, and require a genuine effort to become self-sufficient in
return. Based on the existing literature, it is not clear how we would expect this strategy to
affect the participants’ future employment and earnings prospects. On the one hand, the pro-
gram raises overall benefit levels considerably, and it does so for a relatively long period of
time. There is ample empirical evidence showing that more generous social insurance has
negative effects on labor supply, ceteris paribus; see Krueger and Meyer (2002) for an over-
view of the literature, and Røed and Zhang (2003; 2005), Fevang et al. (2013), and Kostøl and
1 These numbers are computed from administrative register data presented later in this paper.
5
Mogstad (2014) for recent Norwegian evidence. On the other hand, QP requires full-time par-
ticipation in a tailored activation program. There exists no general consensus regarding the
overall impacts of activation. For unemployed job seekers, the typical findings are that there
are favorable “threat effects” prior to active labor market program participation, adverse
“lock-in-effects” during participation, and then (sometimes) favorable “post-program-effects”
afterwards; see, e.g., Kluve et al. (2007) and Card et al. (2010) for recent reviews; and Raaum
and Røed (2006) and Røed and Westlie (2012) for Norwegian evidence. There is also a litera-
ture focusing on programs specifically targeted at temporary disabled persons, but with a few
notable exceptions – Frölich et al. (2004) and Aakvik, et al. (2005), which have indicated
effects close to zero for Sweden and Norway, respectively – this literature is more oriented
toward comparing the effects of alternative rehabilitation strategies than toward evaluating the
more general effects of applying an activation strategy as an alternative to pure income insur-
ance.
The program we evaluate in the present paper is, to our knowledge, unique in its combination
of generosity and (tailored) activation for a hard-to-employ target group with little (or no)
social insurance entitlements. It represents a coherent – yet untested – strategy to fight persis-
tent poverty. Fortunately, the program is also unique in that it was implemented in a way that
facilitates scientific evaluation; i.e., it was phased in gradually over a three-year period, im-
plying that potential participants got access to it at different points in time. In this paper, we
develop a novel evaluation strategy where we combine the staggered program implementation
with observed proxies for “participation propensity” to identify the causal impacts of QP par-
ticipation on subsequent labor market outcomes.
Our main finding is that the program has been successful in terms of helping hard-to-employ
persons into employment. Four years after program entry, we estimate that QP participation
on average raises the employment rate by around 18 percentage points, ceteris paribus. Most
of the extra employment comes in the form of poorly paid and/or very small jobs, however;
hence, the dependency on transfers from the welfare state remains high.
2 Institutions and data
The Qualification Program (QP) was launched in November 2007, and then rolled out gradu-
ally over the next three years –municipality by municipality – in tandem with an administra-
tive reform that merged the local employment and social insurance offices into new joint ad-
6
ministrations (NAV); see, e.g., Christensen et al. (2013). By January 2010, QP had become a
nationwide program. Although the empirical strategy we are going to use in this paper does
not rely on a random-assignment-like order of local implementation, it is worth noting that the
implementation-sequence was determined primarily on the basis of administrative considera-
tions in relation to the establishment of the new NAV-offices, and not on the basis of, say,
local employment opportunities.
QP is designed to support persons who fall between the two stools of employment and social
insurance, and thus potentially face serious poverty problems. The aim of the program is to
provide a stable and safe economic basis over a 1-2 year period, while at the same time help-
ing the participants onto a steady path toward self-sufficiency. The legislation states that “the
Qualification Program applies for persons of working age with substantially reduced work
and income capability and no or very limited social insurance entitlements” (§ 29 in the Law
of Social Services in the Labor and Welfare Administration; our translation). It further states
that entitlement also requires that a) the applicant’s work ability has been assessed, b) that the
program is viewed as both appropriate and necessary in order to increase the applicants’ pos-
sibilities for labor market participation, and c) that the Labor and Welfare administration is
able to offer a suitable program. Given that these (somewhat vague) requirements are met,
access to QP is a legally protected entitlement. The provisions of the Act state that the dura-
tion of the program is to be determined individually. Formally, the QP benefit is granted for
one year at a time, but during the period covered by our analysis, a two-year perspective was
stated as the main rule. Based on an individual assessment the program can also be extended
beyond two years, provided that the claimant has shown progress in his/her efforts to qualify
for the labor market and that a transition to employment appears probable within the near fu-
ture. The program may be terminated at any time if the participant does not fulfill the obliga-
tions set out in the individual plan, or if he/she succeeds in finding regular job.
For the target group of QP, the alternative help offered by the welfare state will typically be
means-tested social assistance or, in cases of serious and lasting health problems, a temporary
or permanent disability insurance benefit. For most of the participants, the program represents
a considerable rise in personal income at the time of entry. Hence, from a pecuniary point of
view, the program is typically viewed as attractive. QP has therefore not been considered use-
ful as a willingness-to-work test whereby caseworkers threaten to terminate social assistance
if participation is rejected. To the contrary, it has been emphasized that participation in QP is
7
voluntary, and should be considered a privilege rather than a duty. In that sense, QP is more a
“carrot” than a “stick”.
Participation in QP entails a fulltime activity of some sort, based on an individually tailored
plan. If the participants nevertheless have additional earnings during the participation period,
the QP benefit is reduced in proportion to the work-hours outside the program, such that, e.g.,
a half-time job results in a 50% reduction in the QP benefit.
Based on reports collected from the municipalities, the Norwegian Labor and Welfare Admin-
istration (NAV) has counted that 17,214 persons had participated in QP by the by the end of
2010, out of which 4,968 had then completed the program according to the individual plan
and 1,414 had dropped out (Schafft and Spjelkavik, 2011). Among those who had completed
the program, 31 % were reported to have obtained regular employment afterwards, whereas 7
% entered regular education. The rest continued receiving some kind of income support, ei-
ther in the form of temporary or permanent disability insurance or in the form of social assis-
tance.
The data we use in the present analysis are collected from administrative registers and com-
prise in principle the whole Norwegian population. Information on individual participation in
QP is based on a separate code for QP benefits used on the paycheck submitted for each par-
ticipant by the municipalities to the tax authorities. This way, we can identify the year of pro-
gram entry and also the years of continued participation, but not the exact starting and stop-
ping dates. By means of encrypted identification numbers, we merge these data to several
other administrative registers containing a wealth of information on individual characteristics
(like gender, age, education, nationality, municipality of residence) as well as longitudinal
information on past and future employment and income sources. From these data, we compute
person-year-observations on several outcomes, particularly related to employment, earnings,
and social insurance dependency.
Based on the procedure described above, we identify 19,211 participants from 2008 through
2011, which is roughly in line with the numbers reported (manually) by the municipalities to
the Labor and Welfare Administration. The number of entrants was largest in 2009 and 2010;
see Table 1. This table also shows that many participants remain in the program for more than
8
two years.2 Figure 1 presents some key descriptive statistics for the participants identified in
our data – in terms of their labor earnings, their employment propensity, and their claims of
(taxable) social insurance benefits and (non-taxable) means-tested social assistance (welfare)
in a period from 8 years before till 3 years after entry into QP. These statistics all indicate that
entry into QP coincided with a marked turning point in economic outcomes for the partici-
pants. Prior to the year of QP entry, the participant group members had experienced a steady
decline in average employment and labor earnings during the whole 8-year period covered by
our data, and a corresponding rise in the dependency on means-tested social assistance. In the
years after entry, these trends were significantly reversed. Moreover, from the year of entry,
the level of taxable benefits also rose markedly, basically reflecting that the QP benefit itself
falls into this category.
Table 1: Descriptive statistics for participants in QP Number of participants 19 211
Entry year
2008 2,919 (15.2 %)
2009 5,857 (30.5 %)
2010 6,060 (31.5 %)
2011 4,375 (22.8 %)
Still participating
1 year after entry year 82.9 %
2 years after entry year 49.8 %
3 years after entry year 23.1 %
Mean age 33.7 years
Female % 44.0 %
Non-native % 50.7 %
High school % 16.1 %
College/University % 7.0 % Note: Statistics are based on all QP entrants from 2008 through 2011. Participation rates 1-3 years after entry are based on entrants than can be followed for the period in question, with data ending in 2011. For example, partic-ipation 3 years after entry is based on 2008-entrants only.
Figure 1 certainly conveys the impression that the QP program must have had large positive
impacts on the participants’ employment and labor earnings trajectories. Note in particular,
that the pre-program decline in labor earnings and employment not at all resembles the noto-
rious “Ashenfelter Dip” (Ashenfelter, 1978) whereby the deterioration occurs just prior to
program entry and, hence, foreshadows a reversion “toward the mean” in the near future re-
gardless of program participation. To the contrary, the pre-program decline appears to have
2 Since we do not have exact starting and stopping dates, we cannot compute accurate durations, but the
numbers in Table 1 indicate that around 50 % of the participants are in the program in at least three consecutive calendar years, and 23 % are participating in at least four years.
9
been a consistent feature of the participants’ economic fortunes for many years, perhaps indi-
cating that the decline would have continued in the absence of program entry.
Figure 1. Annual employment, earnings, and benefit claims among QP participants. Note: The graphs are based on all QP entrants 2008-2011 (19,211 persons); see also note to Table 1. The sizes of the dots are proportional to the number of observations behind each data point. Back-in-time-observations are censored at age 18.
3 Methodology
The research question we seek to answer is how participation in QP affects earnings, em-
ployment, and benefit trajectories for up to four years after the year of program entry; i.e., we
will try to find out how much – if anything – of the patterns displayed in Figure 1 that can be
interpreted as causal. Given the relatively long durations of QP participation, we expect the
causal impacts to change significantly with time since program entry. A probable dynamic
effect pattern is that there are negative (lock-in) effects during the first 1-2 years after entry,
whereas the potentially favorable post-treatment effects build up gradually afterwards. The
statistical models we use to evaluate the program portray a person i who in some base-year t
2040
6080
1000
NO
K
-8 -7 -6 -5 -4 -3 -2 -1 0 +1+2+3+4Years since entry
(a) Labour earnings
.1.1
5.2
.25
.3F
ract
ion
em
plo
yed
-8 -7 -6 -5 -4 -3 -2 -1 0 +1+2+3+4Years since entry
(b) Employment
2040
6080
100
120
1000
NO
K
-8 -7 -6 -5 -4 -3 -2 -1 0 +1+2+3+4Years since entry
(c) Taxable benefits
4050
6070
8090
1000
NO
K
-8 -7 -6 -5 -4 -3 -2 -1 0 +1+2+3+4Years since entry
(d) Non taxable benefits
10
belongs to a risk-group of potential QP entrants during the next four years (t+1,…,t+4) pro-
vided that QP becomes available in person i’s municipality during this period. The model is
then designed to explain various labor market outcomes over a period of up to four years after
the first possible QP entry year (i.e., up to five years after the base-year).3
To define the potential risk-group; i.e., a group of persons who potentially could have partici-
pated in QP – we rely on the eligibility criteria set out in the legislation, and include all per-
sons aged 18-55 who in a base-year t received some kind of temporary income support (ex-
cept sick-pay) from the welfare state, and at the same time had low previous labor earnings
and thus low (or no) social insurance entitlements in the coming years.4 We use the term po-
tential risk-group to emphasize that they are actually at risk only if the program becomes
available in the municipality during the next four years.
The base-years used in our analysis are 2000-2007, with outcomes measured in 2001-2012.
This implies that all the base-years are strictly prior to the first local implementation of the
program, ensuring (by construction) that there is no QP-participation in the base-years. It also
implies that the base-year observations recorded in the first part of our data window (2000-
2003) are never exposed to the risk of actual QP participation in the period we follow them
(since we consider the entry-risk to negligible after four years and since the first municipali-
ties introduced the program in 2008), whereas subsequent base-year cohorts are exposed to an
increasing extent. This pattern is illustrated in Table 2. For example, we see that the 2004
base-year cohort is exposed to QP risk in the fourth year (t+4=2008) provided that they resid-
ed in a municipality implementing the reform in 2008. The only group exposed to QP risk in
the first year after the base-year (i.e., in t+1) is the 2007-cohort in early reform municipalities.
This is also the only group that can be followed for as much as four years after QP entry in
our data.
3 We disregard the risk of QP entry more than four years after the base-year. In principle, we could have modeled entry in a fifth year also, since outcomes are modeled up to five years after the base-year. But, as be-comes clear when we explain our statistical approach, this would have complicated the model considerably without adding anything of substance (the probability of entering QP five years after the base-year is around 0.3 %).
4 The included benefits are unemployment insurance, temporary disability insurance (not including sick-pay, which is payable for a maximum of one year), and social assistance. In the main specification of our model, the definition of low previous labor earnings is that max(last years’ earnings, average earnings last three years) does not exceed 170,000 NOK (measured in 2013-value). By comparison, the average fulltime-equivalent annu-al salary for all employees in Norway in 2013 is approximately 500,000).
11
Table 2. Entry possibilities in four year period after base-year. By base-year and municipality reform year. Reform year 2008
(55 % of population) Reform year 2009
(37 % of population) Reform year 2010
(8 % of population) Base year:
2000 No entry possibility No entry possibility No entry possibility 2001 No entry possibility No entry possibility No entry possibility 2002 No entry possibility No entry possibility No entry possibility 2003 No entry possibility No entry possibility No entry possibility 2004 2008 No entry possibility No entry possibility 2005 2008,2009 2009 No entry possibility 2006 2008,2009,2010 2009,2010 2010 2007 2008,2009,2010,2011 2009,2010,2011 2010,2011
To explain our empirical strategy in more detail, let ,i t ry be a labor market outcome in year
t+r for a base-year observation belonging to year t, and let ,i t r pQP be indicator variables
equal to one for persons who entered the QP program p years before the outcome year in
question (and zero otherwise). Furthermore, let itx be a vector of observed individual charac-
teristics measured in the base-year, including age (44 categories), sex, years of education (8
categories), immigrant status (3 categories), labor earnings (in the base-year and on average
during the three years leading up to the base-year), social insurance benefits, social assistance,
number of months with social assistance, number of months with UI claims, and number of
months with temporary disability insurance benefits. In the absence of unobserved sorting
(i.e., if we are willing to assume that participation is randomly assigned, conditional on ob-
served characteristics), we can regress the various outcomes on a vector of program participa-
tion indicators; e.g.
1
, , ,0
, 2000,...,2007, 1,...,5,r
i t r m t r p i t r p i t rp
y QP u t r
'itx β+ (1)
where m t r is a fixed effect for all combinations of municipality (m), base-year (t) and years
since base-year (r) (with municipality assigned in the base-year) and ,i t ru is a residual.5 Here,
5 We apply this very flexible specification of municipality-time effects throughout the paper, implying
the use of 14,624 dummy variables. Note that this is more general than what we would get from more “standard” municipality-year dummies, since we allow the effects in a given municipality in a given calendar year to depend on time since base-year, i.e., the time since the condition of being dependent on temporary income support was imposed. The reason for this is that the composition of groups that vary with respect to the number of years since they were observed to be in need of temporary income support may be very different; hence they may also re-spond differently to, e.g., fluctuations in local labor market conditions.
12
0 represents the effect on the outcome of having entered the program in the same year,
whereas, e.g., 3 represents the effects of having entered it three years ago; in both cases rela-
tive to not having entered the program at all.
Unfortunately, the assumption of conditional random assignment is unconvincing in this case.
Participation in QP is likely to be highly selective even conditional on a wide range of ob-
served characteristics. Given the character of the program (in particular its explicit targeting
of individuals under high risk of becoming completely reliant on means-tested social assis-
tance), we expect actual participants to be strongly negatively selected; hence, we cannot use
the results from this regression to produce estimates with a causal interpretation. Redesigning
(1) to include individual fixed effects and thus exploit the before-after treatment dimensions
illustrated in Figure 1 for actual QP participants would also not solve the problem, since par-
ticipants’ earnings and employment profiles prior to entry have been anything but fixed, and
since it is probable that QP participation is triggered by unobserved events and/or by changes
in attitudes/motivation that would have broken pre-program outcome-trends even in the ab-
sence of QP.
Instead, we are going to rely on an instrumental variables approach, whereby we use the roll-
out of the program during 2008-2010, interacted with individual predicted participation pro-
pensities, as instruments. Intuitively, identification then relies on the observed relationship
between QP participation propensities and economic performance before/after implementa-
tion of QP in each municipality. A disproportionally positive (negative) development for per-
sons with high QP propensity after actual implementation of QP in their municipality will
then be interpreted as indicative of a positive (negative) effect. This identification strategy
requires that we control for “everything else” that could have induced spurious time trends –
or cross-sectional differences – in the relationship between outcomes and QP participation
propensities. Hence, as we explain in more detail below, our approach is designed to isolate
what we consider to be the reliably exogenous variation arising from the staggered QP intro-
duction only.
A preparatory step in this empirical strategy is to construct the instruments for the four differ-
ent entry alternatives that may become relevant (1, 2, 3, or 4 years after the base-year). We do
this by estimating individual participation propensities as functions of observed explanatory
variables on the basis of the set of baseline-observations for which there is a genuine risk of
entering QP. As it turns out, the selection process into QP appears to have varied quite a lot
13
over time and between municipalities with different reform implementation years (Schafft and
Spjelkavik, 2011). Hence, we cannot simply estimate a single QP participation propensity.
Instead, we estimate separate QP entry probabilities for each relevant combination of base-
year cohort, entry year, and reform-year. As shown in Table 2, this gives us 19 different entry
probabilities to estimate. Let , , ,i t q rz be an indicator variable equal to 1 if person i observed in a
base-year t living in a municipality implementing QP in year q entered the program r years
after the base-year, and let itx be a vector of covariates constituting a restricted/reduced set of
covariates.6 We then estimate the corresponding entry probabilities
, , , ,ˆˆ ( ), 2004,...,2007, 2008,...,2010, 1,...,4i t q r qz f t q r '
it t, rx γ , (2)
for all the (19) existing combinations of t, q, and r in our data using binomial logit models.7
Now, although Equation (2) can be estimated on the subset of actual risk-groups only, the 19
resultant predictions can be attributed to all base-year observations in the dataset (since itx is
always observed). They can then be interpreted as predicted annual hypothetical entry proba-
bilities had the respective combinations of t, q, and r become relevant for i. For further use,
we stack them in a (19 1) vector denoted ˆ itz .
Turning back to Equation (1), we note that what we need in order to instrument the endoge-
nous variables ,i t r pQP is predicted actual QP entry probabilities for the 0-4 years prior to the
outcome year in question. For each outcome year t+r, we construct these such that they repre-
sent the actually corresponding , , ,ˆi t q rz if the program actually was available for the (t,q,r) com-
bination in question, and zero otherwise. Doing this, we end up with five instruments repre-
senting actual entry probabilities relative to the outcome year, which we denote
* * * * * *, , 1 , 2 , 3 , 4it i t r i t r i t r i t r i t rz z z z z z . An example may be clarifying: Consider a per-
son with base year t=2007 who was resident in a municipality implementing the program in
q=2008. Examining a corresponding outcome in 2008 (2007+1), the predicted probability of
6 The variable
itx in Equation (2) contains the same variables as
itx in Equation (1), but with scalar vari-
ables instead of indicator variables. The reason for this is that we need a more restrictive specification in this case to avoid problems with no variation in QP entry for particular values of covariates. The following variables
are included in it
x : Age, sex, education level, immigrant status, earnings in the base-year, max of earnings in the
base-year and in the last three years leading up to the base-year, taxable benefits in the base-year, non-taxable benefits in the base-year, number of months with social assistance in the base-year, number of months with UI benefits in the base-year, and number of months with temporary disability benefits in the base-year.
7 We have also tried out linear probability and probit models, with only minor changes in the results.
14
having entered in the same year is given as *
,2007 1 ,2007,2008,1ˆ
i iz z
, whereas the predicted probabili-
ties of having entered in previous years * * * *
,2007 1 1, ,2007 1 2, ,2007 1 3, ,2007 1 4,( )
i i i iz z z z
are all equal to zero,
since the program did not exist at these points in time. Examining the same person’s outcome
in 2009 (2007+2), we have that the probability of entering in the same year is given by
*
,2007 2 ,2007 ,2008,2ˆ
i iz z
, whereas the probability of having entered last year is *
,2007 2 1 ,2007 ,2008,1ˆ
i iz z
, and
the probabilities corresponding to earlier entries are again zero. Note that
* * * * *
,2007 1 ,2007 2 1 ,2007 3 2 ,2007 4 3 ,2007 5 4i i i i iz z z z z
, since they all refer to entry in the same calendar
year.
Now, given the way we have constructed the five elements in *itz , we could have substituted
these predictions directly for the QP participation indicators in Equation (1). However, we
would then not end up with a correct 2SLS model, since there are a significant number of ac-
tual QP entries that are recorded in municipalities and/or years for which the program does
not exist. The most likely explanation for this is that there are errors in municipality-
assignment or that persons have migrated to another municipality after the base-year. This
represents a potential source of contamination bias (the non-participant group is contaminated
with a number of participants), which unaccounted for will bias estimated effects toward zero.
To avoid this problem, we use the predictions in *itz as instruments for all observed entry deci-
sions instead. A further complication is that there are not one, but four mutually exclusive
endogenous variables in (1) – namely QP entry in the first, second, third, and fourth year after
the base-line year. Although there is a single element in *itz that corresponds directly to each of
these endogenous variables, we include all four elements as instruments for all four endoge-
nous variables. Markussen and Røed (2014) have shown that if there are “cross-effects” im-
plying that each instrument has effects on more than one of the endogenous variables, failure
to account for these cross-effects may yield biased results.8 Hence our first- and second step
linear equations become
*, ,' , qp qp p qp p
i t r p m t r it r i t r pQP ' ' 'it t it q itx β + z σ + d z τ + d z ρ (3)
1
, , ,0
, r
i t r m t r p i t r p i t rp
y QP
' ' 'it t r it q itx β+ +d z τ + d z ρ (4)
8 In our case, the results are almost exactly the same regardless of whether we allow all instruments to
affect all endogenous variables or not. As expected, we also get similar, but somewhat smaller effect estimates if
we use *
itz directly as the first-step predictions.
15
for t=2000,…,2007, r=1,…,5, and p=0,…4, where t rd is a vector of base-year-outcome-year
dummy variables (one dummy for each possible combination of base-year and outcome-year),
qd is a vector of reform-year dummy variables (time constant, but varying across municipali-
ties with different reform implementation years), and ,i t r pQP are the OLS predictions from
(3). Our interest is in the coefficients p , which represents the effects of having entered QP
the same year (p=0), the last year (p=1), and so forth, in all cases relative to non-
participation.
The rationale behind including the control variables ,' 't r it q itd z d z in the statistical model, in
addition to those already included in (1), is as follows:
't r itd z (8 base-years × 5 outcome years × 19 hypothetical entry probabilities = 760 varia-
bles) is included to control for any differences in the outcomes and their time-
developments that correlate systematically with the QP participation propensities.9
'q itd z (3 QP-implementation years × 19 hypothetical entry probabilities = 57 variables) is
included to control for any differences in the correlation between QP propensities and out-
comes between municipalities that implemented the reform at different times.
Hence, by including these controls, we narrow down the variation in participation propensities
used to identify the causal effects to the desired quasi-experimental part of it.
Since the instruments used to identify the causal effects of program participation incorporate
the phasing-in of the program itself, all actual participants have obviously been directly in-
duced to participate by the instruments. In the terminology used by Angrist et al. (1996), all
actual participants are “compliers”, and there exists no “always-takers” or “defiers”. Provided
that QP influences the clients’ outcomes through actual participation in the program only, our
statistical approach thus identifies the average treatment effects among the treated (ATET).
This assumption will be violated if, for example, case workers punish claimants who refuse to
participate in QP (despite that this is not how QP is supposed to be used; see Section 2), or if
welfare-recipients reduce efforts to find work in anticipation of QP participation. Note, how-
ever, that if such side-effects are empirically relevant, our estimates are still valid as measures
of overall impacts of the program measured relative to the number of actual participants, since
9 When participation propensities are estimated by linear probability models, we impose one reference
(zero-restriction) for each of the 19 entry routes to avoid perfect colinearity with it
x .
16
these side-effects also represent causal impacts of being exposed to the program. They can no
longer be interpreted as ATET, however. Yet, from a policy perspective, it is probably the
program’s overall impacts on the target group’s subsequent economic performance, relative to
the program’s overall costs (as determined by the number of actual participants), that is the
parameter of main interest – regardless of how the individual effects are distributed between
actual and potential participants. And that is exactly what comes out of our IV approach.
Our empirical strategy hinges on the assumption that the unobserved composition of the base-
year risk populations was not affected by whether the reform was implemented in 2008, 2009,
or 2010. This assumption could have been violated if, for example, would-be QP participants
migrated to the municipalities with early introduction in order to take advantage of the pro-
gram. Since the last base-year used in our analysis is 2007, and the first municipalities imple-
mented the program in 2008, this does not appear very likely; though not entirely impossible.
The decision to implement QP was taken by the Norwegian Parliament in June 2007, giving
well-informed agents the possibility to move to an early-implementing municipality during
July-December 2007. We can check whether or not such migrations have influenced our esti-
mates by tying persons to the address they had a year before the base-year. We return to this
issue in Section 5 after having presented our main results. Note, however, that endogenous
migrations after the base year (as discussed above) do not invalidate our empirical approach,
since we use initial residency as the basis for our instruments. It makes our instruments weak-
er, however.
A final point to note is that, as we show in the next section, many of the individuals in our
dataset qualify for being included in the risk-group in more than one base-year. Given the way
we construct the analysis data, these persons will contribute with multiple – and sometimes
overlapping – five-year outcome sequences. We have done this to ensure a completely sym-
metric risk-group composition throughout the data window. We take the multiplicity of ob-
servations into account when we compute standard errors, however, by considering all obser-
vations from the same person as a cluster.
4 The analysis population
In this section, we give a brief description of the analysis population used in the statistical
analysis. Given the rather vague eligibility criteria, it is not a trivial exercise to identify the
population at risk of entering QP during a forthcoming four-year period. The formal rules
17
described in Section 2 target persons with substantially reduced labor income capacity and no
or very limited social insurance entitlements. In principle, this implies that everyone who have
had low labor earnings over some time and also received some kind of temporary income
support from the welfare state may become eligible. Since social insurance entitlements in
Norway generally depend on the maximum of labor earnings taken over the last calendar year
and the average over the last three years, we also base our definition of “low labor earnings”
on this maximum. In our main specification, we set the threshold to 170,000 NOK (28,000 $)
(measured in 2013 value). This is roughly one third of the average full-time-full-year earnings
level in Norway. A price we pay for excluding persons above this relatively low earnings-
threshold is that a number of actual participants fall outside the risk-group and are therefore
left out of the analysis. In a robustness analysis, we thus increase the earnings-requirement to
340,000. This raises the overall sample size by 65 % and the number of actual participants
covered by 11 %. Note that we do not exclude persons from the dataset on the basis of the
level of temporary social insurance benefits in the base-year (even though no/low benefit enti-
tlement is among the eligibility criteria), since – given our requirement of low earnings the
last three years – these benefits are likely to become exhausted in the near future.
Table 3. Descriptive statistics analysis population Participants Non-participants Number of base-year observations 21,082 1,386,310
Number of individuals 8,896 307,003
Mean age 32.5 36.7
Women % 61.0 46.8
Non-native % 36.2 15.3
High school % 17.8 35.8
College/University % 6.1 10.8
Mean labor earnings base-year (1000NOK, 2013) 19.5 28.8
Mean social assistance base-year (1000NOK, 2013) 103.4 54.4
Mean taxable benefits (1000NOK, 2013) 50.8 135.4 Note: Averages and fractions are computed over base-year observations. 1000 NOK = 167 $ (based on exchange rate April 2014)
Table 3 shows some descriptive statistics for the sample used in the main part of the analysis.
There are 315,899 individuals included, out of which 8,896 (2.8 %) actually participated in
QP. The participants tend to be quite different from the non-participants, e.g., in the form of
lower labor earnings and taxable social insurance benefits, higher levels of social assistance,
and lower levels of education. Women and immigrants are significantly overrepresented in the
participant group.
18
On average, each individual contributes 4.5 baseline-year observations, and thus 5 4.5 22.5
(partly overlapping) outcome-year observations. A point to note is that our statistical analysis
only includes roughly half of the 19,211 QP participants described in Section 2 above. The
main reason for this is that to ensure a completely symmetric composition of analysis popula-
tions in municipalities with different QP implementation dates, our risk groups are defined on
the basis of individual characteristics strictly prior to program implementation; i.e., no later
than 2007. As a result, we lose persons who entered the risk group for the first time during
2008-2011, and this alone accounts for 62 % of the overall loss of actual participants in the
analysis data. In addition, our definition of the risk population is imperfect, implying that
some persons enter QP even though they were not considered (by us) to be at risk; i.e., they
had too high income in the base-year, were too old, or did not receive the types of temporary
income support that we have used to define the risk population. The latter is particularly rele-
vant for humanitarian immigrants who sometimes enter the program directly without first
receiving the temporary transfers. As a result, the participant group included in our statistical
analysis deviates somewhat from the group of all participants described in Table 1. In particu-
lar, we oversample female and undersample immigrant participants.
5 Results
We start out this section by summarizing the computation of the instruments and the results
from the first-step equations. We thereafter turn to the estimated effects of actual participa-
tion.
5.1 Explaining QP participation: The instruments and the first-step equations
Given that we estimate as much as 19 different participation models based on Equation (2)
(depending on base-year, the number of years that have passed since the base year, and the
timing of the reform in the municipality of residence) and the relatively large number of ex-
planatory variables included in each model, we do not present these results in any detail. In-
stead, to illustrate the identified participation patterns, we compute the average of the 19
probability-predictions for each individual base-year observation, and show how the distribu-
tions of the resultant participation propensity summary statistics vary between participants
and non-participants, and how they correlate with the most important observed characteristics
used in the estimation process. As it turns out, the average estimated annual QP entry proba-
bilities (taken over the 19 possible entry-routes) are low for virtually everyone; 3.5 % for par-
ticipants and 1.2 % for non-participants. Over a four-year risk period, this implies that actual
19
participants according to our model had a participation probability around 13.3 %, whereas
non-participants had a participation probability around 4.7 %. The relatively low participation
probabilities estimated even for participants reflects that it is difficult to identify a genuine
high-risk groups based on observed characteristics only, and thus highlights the magnitude of
the selection problem. Nevertheless, observed characteristic are associated with significant
differences in participation propensities. This is illustrated in Figure 2, where we plot the cu-
mulative distribution functions for the predicted participation propensities separately for par-
ticipants and non-participants.
Figure 2. The cumulative distribution of predicted average annual QP participation proba-bilities Note: The distribution functions are based on the averages of the 19 hypothetical QP propensities predicted from Equation (2). Figure 3 further illustrates important aspects of the sorting process by plotting the mean val-
ues of various observed characteristics by percentile in the distribution of average QP partici-
pation probabilities. Those with the highest participation probabilities are young immigrants
with almost zero labor earnings and little schooling, who have virtually no access to (taxable)
social insurance transfers, and relatively large amounts of means-tested social assistance. At
the other end, those with the lowest participation probabilities are older natives, with some
previous labor earnings and schooling, significant social insurance transfers, but virtually no
means-tested social assistance.
00.
250.
500.
751
5% 10% 15%Predicted annual participation propensity
Non-participants Participants
20
Figure 3. Mean values of selected individual characteristics by percentile in the distribution of average predicted QP participation propensity. Note: QP participation propensity is computed as the average of the 19 hypothetical QP propensities predicted from Equation (2).
Table 4 presents the estimation results from the first stage (Equation (3)). Unsurprisingly, we
find strong effects of predicted participation on actual participation. Note, however that the
diagonal elements of Table 4 are not equal to one, nor are the off-diagonal elements always
equal to zero. These patterns primarily reflect that participation sometimes occur in munici-
palities/years that were not considered to be at risk when the instrument was constructed.
Moreover, while the instruments are constructed as a non-linear function of individual charac-
teristics, they enter linearly in the IV model.
3035
4045
50Y
ear
s
0 .05 .1Participation propensity
(a) Age
0.2
.4.6
.81
%
0 .05 .1Participation propensity
(b) Non-native0
1020
3040
5010
00 N
OK
(2
013)
0 .05 .1Participation propensity
(c) Labor earnings
68
1012
14Y
ear
s0 .05 .1
Participation propensity
(d) Years of schooling
010
020
030
040
010
00 N
OK
(2
013)
0 .05 .1Participation propensity
(e) Taxable benefits
050
1001
502
002
50
1000
NO
K (
201
3)
0 .05 .1Participation propensity
(f) Non-taxable benefits
21
Table 4. First stage estimates (Eq. 3), based on sample with low earnings threshold. Effects of excluded instruments on entry probabilities in year t+r-p for p=0,1,2,3,4 (standard errors in parentheses)
, 0i t rQP , 1i t rQP , 2i t rQP , 3i t rQP , 4i t rQP
*, 0i t rz 0.740***
(0.025) -0.015 (0.012)
-0.002 (0.006)
-0.004 (0.003)
0.002 (0.001)
*, 1i t rz 0.011
(0.030) 0.740*** (0.026)
-0.014 (0.013)
0.002 (0.007)
-0.006** (0.003)
*, 2i t rz -0.056**
(0.027) 0.009
(0.036) 0.741*** (0.028)
-0.013 (0.014)
0.010* (0.006)
*, 3i t rz 0.020
(0.023) -0.074** (0.038)
0.014 (0.040)
0.752*** (0.030)
-0.023** (0.011)
*, 4i t rz -0.031*
(0.018) 0.023
(0.038) -0.075 (0.046)
0.037 (0.047)
0.796*** (0.035)
F-statistic excluded instruments
15.82 15.82 18.69 21.95 25.21
Note: Standard errors are clustered at individuals. *(**)(***) indicate significance at the 10(5)(1) percent levels. Number of observations: 7,036,980. Control variables: Municipality×base-year×years-since-base-year-fixed effects (18,280 dummy variables), age (44 dummy variables), sex, education (8 dummy variables), immigrant status (3 dummy variables), labor earn-ings (in the base-year and as average over the three years leading up to the base-year), taxable benefits in the base-year, non-taxable benefits in the base-year, number of months with social assistance in the base-year, num-ber of months with UI benefits in base-year, number of months with temporary disability benefits in base-year, estimated QP participation propensities interacted with base-year×outcome-year (760 variables), and estimated QP participation propensities interacted with reform-year in the municipality (57 variables).
5.2 The causal effects of QP participation: The second-step equations
We are interested in the effects of QP participation on a number of outcomes. Since the main
aim of the program is to help persons into regular employment, we focus on employment as
the main success criterion. For each year, employment is defined as having labor-related earn-
ings (from unsubsidized work) above certain thresholds, which we set relatively low given the
program’s target group of hard-to-employ persons. In addition, we use outcomes measuring
the level of earnings as well as the level of various types of welfare state transfers. For the
latter outcomes, we of course have considerable a priori knowledge about the true causal ef-
fects of the QP program, since the program by design offers a taxable full-year-equivalent
transfer of 170,000 NOK (28,000 $), which to some extent substitutes for non-taxable benefits
(means-tested social assistance). This implies that we can use the models’ estimated effects on
these outcomes as a sort of plausibility-test.
We present our main results regarding employment and earnings in Table 5. For comparison,
we present the results from our preferred IV specification (based on Equation (4)) together
with the “naïve” OLS-estimates (based on Equation 1). While the OLS estimates consistently
indicate adverse effects of QP in the form of lower employment and lower earnings, the IV-
estimates indicate favorable effects from 2-3 years after entry. In particular, QP participation
seems to raise the chances of obtaining at least some employment quite significantly. Based
22
on the widest employment definition (annual earnings exceeding at least 85,000 NOK), the
point estimates indicate a 10 percentage point increase in the employment probability two
years after QP entry, a 14 point increase after three years, and an 18 point increase after four
years (Column II). The estimated impacts on the levels of annual earnings follow a similar
time pattern (Column IV); after four years, QP participation is predicted to raise annual earn-
ings by around 50,000 NOK (8,300 $). Given that the average fulltime-equivalent annual
earnings level in Norway is around 500,000 NOK, this is in quantitative terms not a huge ef-
fect. However, the linear earnings equation is poorly designed to capture the true earnings
effects of QP, since the effects are likely to be concentrated in the extreme lower tail of the
earnings distribution. Using a log-specification instead (Column XII), we estimate a QP im-
pact of 1.4 log-points after four years, indicating that those who actually participate in the
program (and typically have very poor earnings prospects) experience large relative increases
as a result of QP participation.
Although the statistical uncertainty associated with all the employment and earnings estimates
is relatively large when evaluated one by one, it is notable that the overall pattern of estimated
coefficients conveys a coherent and plausible story: QP participation reduces employment
and labor earnings slightly during the first 1-2 years of participation and raises them after-
wards. At the same time it sharply (and statistically significantly) increases the level of taxa-
ble benefits (Column VIII) and reduces the level of social assistance (Column X). The sizes of
the latter effect-estimates correspond closely to what we would expect on the basis of prior
knowledge. For example, the estimated rise in taxable benefits of 106,000 NOK in the first
year after QP entry accords well with the fact that annual full-year benefits are equal to
170,000 NOK, as some QP entrants would have been eligible for small amounts of taxable
benefits even without the program and as some entrants do not participate the whole year.
As pointed out in Section 3, provided that QP affects actual participants only, our effect esti-
mates can be interpreted as the estimated average treatment effect among the treated (ATET).
If there are indirect effects on non-participants, e.g., in the form of “threat effects” or in the
form of changes in the local treatment culture that spill over to other social assistance recipi-
ents, our IV estimates are still valid as measures of the overall effects relative to the number
of treated. But, since they attribute all (reduced form) effects to the actual participants they
will in this case underestimate the number of affected clients and thus overestimate the causal
effect for each of them.
Table 5. Main estimation results (Eq. 4) for sample with low earnings threshold. Ordinary Least Squares and Instrumental Variables (clustered standard errors in pa-rentheses)
Employment
(85,000 NOK thresh-old)
Employment (170,000 NOK thresh-
old)
Labor earnings (1000 NOK, 2013-value)
Taxable benefits (1000 NOK, 2013-
value)
Non-taxable benefits (1000 NOK, 2013
value)
Log labor earnings (log( earnings+1))
(1000 NOK, 2013 value) I II III IV V VI VII VIII IX X XI XII
Eq. 1 (OLS)
Eq. 4 (IV)
Eq. 1 (OLS)
Eq. 4 (IV)
Eq. 1 (OLS)
Eq. 4 (IV)
Eq. 1 (OLS)
Eq. 4 (IV)
Eq. 1 (OLS)
Eq. 4 (IV)
Eq. 1 (OLS)
Eq. 4 (IV)
Effects of QP
Same year (p=0) -0.201***
(0.003) -0.086 (0.074)
-0.177*** (0.002)
-0.047 (0.065)
-60.02*** (0.66)
-26.33 (21.40)
12.70*** (0.68)
43.50*** (15.66)
45.18*** (0.68)
-27.00* (13.92)
-0.829*** (0.019)
-0.133 (0.396)
First year after entry (p=1)
-0.145*** (0.004)
0.010 (0.064)
-0.128*** (0.003)
-0.061 (0.058)
-47.00*** (0.88)
-28.18 (19.11)
55.41*** (0.83)
106.50*** (17.86)
21.57*** (0.72)
-88.12*** (13.33)
-0.735*** (0.021)
-0.267 (0.330)
Second year after entry (p=2)
-0.075*** (0.005)
0.096 (0.079)
-0.071*** (0.004)
0.028 (0.071)
-28.35*** (1.20)
4.25 (24.37)
30.01*** (1.05)
-10.92 (19.25)
11.59*** (0.81)
-16.01 (17.02)
-0.448*** (0.026)
0.304 (0.431)
Third year after entry (p=3)
-0.042*** (0.006)
0.143* (0.082)
-0.034*** (0.005)
0.081 (0.073)
-15.34*** (1.64)
12.18 (24.96)
19.86*** (1.31)
-25.23 (20.93)
0.54 (0.94)
-45.84*** (16.09)
-0.261*** (0.033)
0.465 (0.434)
Fourth year after entry (p=4)
-0.012 (-0.010)
0.182** (0.089)
-0.008 (0.009)
0.121 (0.079)
-5.80** (2.72)
50.54* (26.71)
23.53*** (2.12)
-32.08 (22.10)
-8.23*** (1.58)
6.73 (16.90)
-0.096* (0.053)
1.395*** (0.477)
Note: Standard errors are clustered at individuals. *(**)(***) indicate significance at the 10(5)(1) percent levels. Number of observations in all models: 7,036,980. 1000 NOK = 167 $ (based on exchange rate April 2014). Control variables included in all models: Municipality×base-year×years-since-base-year-fixed effects (18,280 dummy variables), age (44 dummy variables), sex, education (8 dummy variables), immigrant status (3 dummy variables), labor earnings (in the base-year and as average over the three years leading up to the base-year), taxable benefits in the base-year, non-taxable benefits in the base-year, number of months with social assistance in the base-year, number of months with UI benefits in base-year, and number of months with temporary disability benefits in base-year. Additional control variables in IV-models: Estimated QP participation propensities interacted with base-year×outcome-year (760 variables), and estimated QP participation propensities interacted with reform-year in the municipality (57 variables).
Table 6. Estimation results for sample with high earnings threshold. Instrumental Variables estimates (Eq. 4) (clustered standard errors in parentheses) Whole population Natives only Immigrants only I II III IV V VI VII VIII IX Employment
(>85,000 NOK)
Employment (>170,000
NOK)
Log labor earnings
Employment (>85,000
NOK)
Employment (>170,000
NOK)
Log labor earnings
Employment (>85,000
NOK)
Employment (>170,000
NOK)
Log labor earnings
Effects of QP
Same year (p=0) -0.043 (0.069)
0.017 (0.062)
0.055 (0.372)
-0.054 (0.079)
-0.023 (0.069)
0.210 (0.471)
0.017
(0.118) 0.080
(0.110) 0.193
(0.624)
First year after entry (p=1)
0.052 (0.060)
-0.008 (0.056)
-0.115 (0.309)
0.028
(0.070) -0.007 (0.064)
-0.307 (0.407)
0.018
(0.104) -0.073 (0.100)
-0.107 (0.520)
Second year after entry (p=2)
0.156** (0.074)
0.138** (0.067)
0.668* (0.402)
0.139* (0.083)
0.125* (0.074)
0.808 (0.499)
0.238
(0.146) 0.211
(0.138) 1.124
(0.757)
Third year after entry (p=3)
0.198** (0.077)
0.131* (0.071)
0.629 (0.408)
0.201** (0.085)
0.105 (0.077)
0.718 (0.512)
0.001
(0.181) -0.094 (0.175)
-0.424 (0.893)
Fourth year after entry (p=4)
0.273*** (0.084)
0.212*** (0.076)
1.848*** (0.449)
0.219** (0.091)
0.132* (0.080)
2.046*** (0.576)
0.404
(0.356) 0.473* (0.254)
2.747* (1.267)
Note: Standard errors are clustered at individuals. *(**)(***) indicate significance at the 10(5)(1) percent levels. 1000 NOK = 167 $ (based on exchange rate April 2014) Number of observations: Column 1-3: 11,567,275; Column IV-VI: 9,953,115; Column VII-IX: 1,451,965. Control variables included in all models: See note to Table 5.
5.3 Robustness across samples and subgroups
How robust are these results? In this sub-section, we look at how the key estimates change
when we expand the analysis population to include persons with somewhat higher earnings in
the base year (or in the two years prior to the base year), as discussed in Section 4. In addi-
tion, given the large fraction of immigrants participating in the program, we estimate separate
models for natives and immigrants. In these exercises, we focus on the outcomes of employ-
ment and log labor earnings only.
We first re-estimate the model based on a much wider definition of the risk-population in-
creasing the sample size by around 65 %. This is done by raising the upper earnings threshold
in the base-year to NOK 340,000 (57,000 $). The results from this exercise are presented in
Table 6, Columns I-III. The results are similar to those reported in Table 5. The main differ-
ence is that the estimated employment and earnings effects 2-4 years after QP entry become a
bit larger and also more statistically significant. Based on the larger sample, we also estimate
separate models for natives and immigrants. The results for natives are in line with the results
reported for the complete population; see Table 6, Columns IV-VI. The results for immigrants
also display a similar pattern, although they are much more unstable and subject to statistical
uncertainty. Hence, we do not really have sufficient information in the data to draw any firm
conclusions regarding possible effect heterogeneity.
5.4 Endogenous migration?
As discussed in Section 3, a causal interpretation of our estimates hinges on the assumption
that persons’ addresses (municipalities) up to (and including) 2007 are exogenous. Since the
QP program was legislated in June 2007, there was a short window of opportunity during the
autumn of 2007 whereby would-be QP participants could self-select into municipalities with
early (2008) QP implementation. On average, around 11 % the risk-group population move
across municipalities each year, but we see no pattern of increased migration during the peri-
od in question. To nevertheless check the sensitivity of our estimates with respect to endoge-
nous migration, we re-estimate the models presented in Section 5.2, but this time we tie each
base-year observation to the last years’ (t-1) municipality (i.e., we disregard migrations that
occurred between t-1 and t). This way, we eliminate the possibility of an endogeneity prob-
lem, but at the cost of inducing more measurement error into the model. The results of this
exercise are presented in Table 7. As it turns out, the estimates change very little, suggesting
that endogenous migration is not a big issue in the present context.
26
Table 7. Estimation results with last year’s address used to identify municipality (dataset with low earnings threshold). Instrumental Variables estimates (Eq. 4) (clustered standard errors in parenthe-ses) I II III
Employment
(>85,000 NOK) Employment
(>170,000 NOK) Log labor earn-
ings Effects of QP
Same year (p=0) -0.050 (0.080)
-0.025 (0.070)
0.102 (0.432)
First year after entry (p=1) 0.024
(0.068) -0.066 (0.063)
-0.252 (0.354)
Second year after entry (p=2) 0.139
(0.087) 0.065
(0.078) 0.604
(0.473)
Third year after entry (p=3) 0.158* (0.088)
0.086 (0.080)
0.341 (0.471)
Fourth year after entry (p=4) 0.194* (0.099)
0.100 (0.088)
1.348** (0.533)
Note: Standard errors are clustered at individuals. *(**)(***) indicate significance at the 10(5)(1) percent levels. 1000 NOK = 167 $ (based on exchange rate April 2014). Number of observations: 6,903,650. Control variables included in all models: See note to Table 5.
5.5 Benefits versus costs
Our analysis indicates that QP to some extent has accomplished its aims of helping hard-to-
employ persons back to (or into) work. However, the program has also been costly. Have the
benefits outweighed the costs? Based on the admittedly highly uncertain point estimates from
the baseline model (Table 5), we calculate that the overall increase in net (after tax) transfers
generated in the QP participation period (defined as the entry year plus the two subsequent
years) amounts to approximately 50,000 NOK (8,300 $). In addition, the increased adminis-
trative costs associated with fewer clients per caseworker (see the introduction) also amount
to around 50,000 NOK. Adding a 20 % marginal cost of public funds and leaving out the pure
transfer component, these numbers imply a direct social cost of 70,000 NOK (11,700 $) for
each participant. Moreover, labor earnings are estimated to decline by approximately 55,000
NOK cumulatively over the first two years (including the entry year), as a result of lock-in
effects. On the benefit side, labor earnings start to increase from the second year after entry,
and reach a positive annual QP effect of 50,000 NOK after four years. However, the cumula-
tive earnings gain in this period (67,000 NOK) still falls short of the initial costs. Hence, for
this particular cost-benefit assessment to come out with a positive number, we need to assume
that the favorable earnings and employment effects to some extent persist after the fourth
year. If we do extrapolate the estimated forth-year effects to subsequent years, our highly
simplified cost-benefit calculation breaks even in the fifth year, and becomes positive thereaf-
ter.
27
In addition to the costs and benefits discussed here, there will of course also be costs and ben-
efits associated with the particular activities that the QP participants take part in, which – as
described above – ranges from medical rehabilitation to fulltime work. Since we do not have
any specific information about individual QP activities, nor about the activities that would
have prevailed without QP, we are not able to provide precise assessment of these cost-benefit
components. However, statistics reported by Statistics Norway, based on information collect-
ed from the municipalities, indicate that work is the dominating activity. For example, among
those who entered QP in 2009, as much as 78 % participated in some form of employment in
2010 as part of their individual QP plan.10 This presumably implies that value is generated,
though probably at some costs associated with individual support and workplace adaptations.
Additional favorable effects of QP may come from peer influences. Recent empirical studies
have indicated that welfare dependency is contagious within social networks; see, e.g., Rege
et al. (2012) and Markussen and Røed (forthcoming). Hence, if QP succeeds in moving at
least some persons from welfare to work, we may expect to see subsequent knock-on effects
among their peers.
An important point to bear in mind is that a main aim of QP was to reduce poverty. Our esti-
mation results indicate that the overall level of take-home earnings was raised for all years
after entry, with a possible exception for the third year after entry in which the increased
transfer level has been tapered off while the favorable earnings effects are still small. Hence,
poverty problems appear to have been alleviated by the program.
6 Conclusion
The main conclusion of our analysis is that the combination of activation requirements and
economic generosity appears to have had the intended effect of helping a very hard-to-employ
group of persons back to (or into) work in Norway. Two years after entry into the Qualifica-
tion Program (QP), we identify indications of a positive effect on the probability of having
obtained at least some paid employment, and the effect increases in subsequent years. Four
years after entry, our estimates indicate that the employment probability is raised by as much
as 18 percentage points, ceteris paribus. Effect estimates of this order appear robust across
different samples and subgroups. The statistical uncertainty associated with each single esti-
10 These numbers are available from «Statistikkbanken» (https://www.ssb.no/statistikkbanken).
28
mate is large, however (standard errors around 8-9 percentage points), hence the point esti-
mates should be interpreted with care.
Although most of the additional jobs are small, at least to start with, they may represent im-
portant stepping-stones into regular employment for a group of persons who in the absence of
the program most likely would have had labor force participation rates close to zero. Together
with the structured and active everyday life implied by the program itself this may also have
built up a work habit among the participants with possible knock-on effects on the partici-
pants’ peers. Whether the gains identified in this paper are sufficient to compensate for the
costs depends on the extent to which the moderately increased labor earnings can be expected
to last or perhaps even become a first step toward full employment, which is too early to say
based on our data.
References
Aakvik, A., Heckman, J., and Vytlacil, E. (2005) Estimating Treatment Effects for Discrete Out-
comes when Responses to Treatment Vary: An Application to Norwegian Vocational Reha-
bilitation Programs. Journal of Econometrics, Vol. 125, 15-51.
Angrist, J., Imbens, G. W. and Rubin, D. B. (1996) Identification of Causal Effects Using Instrumen-
tal Variables. Journal of the American Statistical Association, Vol. 91, 444-455.
Ashenfelter, O. (1978) Estimating the Effect of Training Programs on Earnings. Review of Economics
and Statistics, Vol. 60, 47-50.
Calvó-Armengol, A. and Jackson, M. O. (2004) The Effects of Social Networks on Employment and
Inequality. American Economic Review, Vol. 94, No. 3, 426-454
Card, D., Kluve, J., and Weber, A. (2010) Active Labour Market Policy Evaluations: A Meta-
analysis. The Economic Journal, Vol. 120, F452–F477.
Christensen, T., Fimreite, A. L., and Lægreid, P. (2013) Joined-Up Government for Welfare Admin-
istration Reform in Norway. Public Organization Review, forthcoming.
Dasgupta, P. (1997) Nutritional Status, the Capacity for Work, and Poverty Traps. Journal of Econo-
metrics, Vol. 77, 5-37.
Dasgupta, P. and Ray, D. (1986) Inequality as a Determinant of Malnutrition and Unemployment:
Theory, Economic Journal, Vol. 96, 1011-1034.
Dasgupta, P. and Ray, D. (1987) Inequality as a Determinant of Malnutrition and Unemployment:
Policy, Economic Journal, Vol. 97, 177-188.
Fevang, E., Hardoy, I., and Røed, K. (2013) Getting Disabled Workers Back to Work: How Important
Are Economic Incentives? IZA Discussion Paper No. 7137.
29
Kostøl, A. R., and Mogstad, M. (2014) How Financial Incentives Induce Disability In-surance Recip-
ients to Return to Work. American Economic Review, forthcoming.
Kluve, J., Card, D., Fertig, M., Góra, M., Jacobi, J., Jensen, P., Leetmaa, R., Nima, L., Patacchini, E.,
Schaffner, S., Schmid, C. M., van der Klaauw, B., and Weber, A. (2007) Active Labor Market
Policies in Europe - Performance and Perspectives. Springer-Verlag Berlin and Heidelberg
GmbH & Co.
Frölich, M., Hesmati, A., and Lechner, M. (2004) A Microeconomic Evaluation of Rehabilitation of
Long-term Sickness in Sweden. Journal of Applied Econometrics, Vol. 19, 375-396.
Gubrium, E. K. and Lødemel, I. (2014) “Not Good Enough”: Social Assistance and Shaming in Nor-
way. Chapter 5 in Gubrium, E. K., Pellissery, S., and Lødemel, I. (Eds.): The Shame of It.
Global Perspectives on Anti-Poverty Policies. Policy Press. Bristol, UK.
Krueger, A. B. and Meyer, B. D. (2002) Labor Supply Effects of Social Insurance. In Auerbach, A. J.
and Feldstein, M. (eds.) Handbook of Public Economics, Vol. 4. Elsevier Science, North-
Holland, 2002; 2327-92.
Markussen, S. and Røed, K. (forthcoming) Social Insurance Networks. Journal of Human Resources.
Markussen, S. and Røed, K. (2014) The Impacts of Vocational Rehabilitation. IZA Discussion Paper
No. 7892.
Moffitt, R. (2007) Welfare Reform: The US Experience. Swedish Economic Policy Review, Vol. 14,
No. 2, 11-48.
Mullainathan, S. and Shafir, E. (2013) Scarcity: Why Having Too Little Means so Much. Time Books,
Henry Holt and Company, New York.
Rege, M., Telle, K., and Votruba, M. (2012) Social Interaction Effects in Disability Pension Partici-
pation – Evidence from Plant Downsizing. Scandinavian Journal of Economics Vol. 114, No.
4, 1208–1239.
Røed, K. (2012) Active Social Insurance. IZA Journal of Labor Policy, Vol. 1:8 (doi:10.1186/2193-
9004-1-8).
Røed, K. and Raaum, O. (2006) Do Labour Market Programmes Speed up the Return to Work? Ox-
ford Bulletin of Economics & Statistics , Vol. 68, No. 5, 541-68
Røed, K. and Westlie, L. (2012) Unemployment Insurance in Welfare States: The Impacts of Soft
Duration Constraints. Journal of the European Economic Association, Vol. 10, No. 3, 518-
554.
Røed, K. and Zhang, T. (2003) Does Unemployment Compensation Affect Unemployment Duration?
Economic Journal, Vol. 113, 190-206.
Røed, K. and Zhang, T. (2005) Unemployment duration and economic incentives - A quasi random-
assignment approach. European Economic Review , Vol. 49, 1799-1825.
Schafft, A. and Spjelkavik (2011) Evaluering av Kvalifiseringsprogrammet (Evaluation of the Quali-
30
fication Program). AFI-rapport 4/2012.
Shah, A. K., Mullainathan, S., and Sharif, E. (2012) Some Consequences of Having Too Little. Sci-
ence, Vol. 338, 682-685,
Walker, R., Kyomuhendo, G. B., Chase, E., Choudhry, S., Gubrium, E. K, Nicola, J. Y., Lødemel, I.,
Mathew, L. Mwiine, A., Pellissery, S. and Ming, Y. (2013) Poverty in Global Perspective: Is
Shame a Common Denominator? Journal of Social Policy, Vol. 42, No. 2, 215-233.